Computers Only Do One Thing
In March of this year, music industry giant Warner Music Group signed a 20 album deal with Endel, a startup which uses its own custom software to generate music.
While some people reacted with interest, there was a noticeable tone of unease in many of the responses. I even saw expressions of sadness, anger, or disgust. Articles were written with titles like “The End is Nigh: An algorithm just signed with a major record label.”
Some expressed fears of music-generation software replacing human artists, a thought which was prevalent enough to elicit a response from Endel:
More recently, the company OpenAI released software called MuseNet, “a deep neural network that can generate 4-minute musical compositions with 10 different instruments.” Technology Review covered the announcement with a dismissive article which questions the creative abilities of AI and goes so far as to ask if MuseNet’s output is “really music.”
This AI-generated musak shows us the limit of artificial creativity
If Mozart were alive today (and if he was feeling a bit uninspired) he might well sit down and produce a piece of music…
I feel there is a pervasive attitude that art, especially music, which is created by a computer is degenerate. Many people tend to regard computer-generated art as “beneath” that which was directly created by humans. Here’s a video of legendary anime director Hayao Miyazaki in which a research team explains their intention to create software which “can draw pictures like humans do.” Miyazaki reacts dramatically, stating, “I feel like we are nearing to the end of times. We humans are losing faith in ourselves.”
I agree with those who value human creativity, and I empathize with their fears of a dystopian, artist-less future. However, I believe most aversions to computer-generated art are predicated on a misunderstanding of our relationship with technology.
The Only Thing Computers Do
The reality is that computers can really only do one single thing: what we tell them to.
Okay, so it’s a bit clickbaity. Obviously, computers can do lots of things. However, all of the things computers do, they do only because we told them to. We told computers to file our taxes. We told computers to talk to each other. We told computers to drive cars.
Computers, and the software they execute, are just tools. Like hammers, paintbrushes, guitars, and cameras, they are worthless without humans to utilize them. Anything we want a computer to do must be programmed, and no computer has ever done even one single thing it wasn’t programmed to do.
Computers don’t spontaneously decide to produce art — computers don’t even know what art is. They only produce art if humans program them to, and this process requires the programmer to define the possibilities of the program’s execution. For example, the act of programming a computer to produce images necessarily means the programmer must define what kind of images are produced. Does it produce one single image or can it produce more? What sizes and shapes of images can it produce? What colors can it use? Does it make abstract images or does it draw figures? Are the images photo-realistic or are they more stylistic? What styles of images can it produce? Every aspect of the potential output of an art-producing program is both defined and controlled by the programmer.
Taste and the Creative Process
How do artists make decisions while creating art? They use their artistic taste.
Imagine you and I are asked to explain how we feel about a portrait and why we feel that way. Of course, one of us might like the portrait while the other does not, but we might even disagree about which aspects of the painting we considered when making our judgments. Perhaps you like the emotion displayed in the face of the painting’s subject, but I either didn’t notice or don’t care. Even if we agree that the colors of the painting influence our opinion of it, we might still disagree as to whether or not we like them. Neither of us can be right, as we could never justify to each other why the metrics we used are the appropriate metrics for the painting, nor can we establish a way to objectively measure those metrics. We have completely different tastes.
When artists create works of art, they use their tastes to make decisions about the works. Taste dictates the words used by a lyricist, the lighting used by a photographer, and the shapes molded by a sculptor.
Likewise, when programming a computer to create art, the programmer must necessarily use their taste to dictate the output. In this way, the programmer is the artist of the work, and the computer is simply a tool used to produce it.
The music created by Endel did not magically appear on its own. It is generated by software built by humans, and it’s created exactly according to their specifications. These specifications are selected based on the tastes of people at Endel. In fact, the company’s co-founders and software engineers are credited as songwriters on the albums released through WMG, which I consider perfectly appropriate.
Artificial Intelligence and Machine Learning
While AI/ML techniques may at first seem to exclude the programmer’s taste from the creation process, this is not accurate. Again, programs which use these techniques are just tools, and they must be applied by humans to create anything. While the purpose of these techniques is indeed to create output which is similar to some preexisting input, the ways these techniques are used and the interpretation of the inputs are still dictated by the programmer’s taste.
For a project like MuseNet, consider how many decisions are made based on the programmers’ tastes. Is the order of the notes the most important aspect to consider? What about the duration of each note? What about harmony? What qualifies as a melody? Which of these aspects should be considered, and which should not? Should a LSTM network be used or a different type of neural network? Which inputs should be used to train the model, and which should not? How do you even measure the similarity of the input to the output? Answering these questions requires programmers to reveal their own opinions about the output and imbue the program with their tastes.
The primary misunderstanding with AI/ML-based art generation techniques is thinking of the program as the artist, rather than as the tool. This is exemplified in the Technology Review article mentioned earlier by a statement in the last paragraph:
“…it’s unclear whether AI can be creative at all.”
Allow me to illustrate the problem with that sentence:
It’s unclear whether paintbrushes can be creative at all.
Origins of Misunderstanding
Why do people tend to view computers as a replacement for artists, rather than as a tool used by them? I can think of three possible factors:
- The creation process is less relatable.
- The work is indirectly created by the artist.
- Decisions about the work are made by the computer.
We understand paintings were painted by a human; many of us have seen painters work or even tried it ourselves. When we see music performed live, we’re watching artists make art right in front of us, and when we listen to recorded music, we know the singer and instrumentalists had to actually perform what we’re hearing. Most of us have made music before, either by playing an instrument or even just singing in the shower.
The tools used for painting, making music, and creating other art have been available for a very long time, and they’re accessible to everyone. Comparatively, computers are young, and significantly fewer people understand how to program. It’s easy to forget the music generated by Endel was programmed by humans because we didn’t see it happen, and even if we did, most of us wouldn’t understand it anyway.
Here is a well-known video of Wintergatan’s Marble Machine. Much like a computer running Endel’s software, it’s a system which produces music.
The Marble Machine video has had an overwhelmingly positive response. Of course, a very notable difference is the Marble Machine is directly operated by a person, and we can observe the creation process; it just seems like a big, complicated instrument. I wonder, if the human operator of the Marble Machine were replaced by a machine, would there be any question as to who was the artist of the work? Probably not, as we can see the workings of the machine, and we understand that it was built by a human. Contrast this with software, which to most people is basically just a black box of mystery.
Another likely contributor to the misunderstanding of computer-generated art is the indirect nature of the creation process. Rather than creating the final output, computer artists build systems which then produce some output. This is unlike most other art-creation processes, where artists directly create works of art. When we see a computer artist step back as their system creates art, it gives the illusion that the programmer hasn’t done anything at all, and is simply in possession of a magic art-creation machine. Once the software starts producing art “at the push of a button,” it’s easy to forget the programmer has complete control of the output.
Finally, when a computer makes decisions about a work of art, it can seem as though the program should get some of the artistic credit. However, as mentioned before, it’s the programmer who must necessarily specify the range of possible outcomes of every decision made by the computer. Computers merely follow instructions, and they don’t have opinions about a decision’s options. Unfortunately, this isn’t always obvious to an observer of computer-generated art, which once again creates the illusion that the computer is the artist.
If you run the following code snippet, you’ll just see a single color:
Which color do you see? You might see pink, or cyan, or “coral” (an orangey color). I don’t know which one you see, because I programmed your computer to randomly choose one of those three colors. On the surface, it might appear as though the computer made an artistic choice about which color to show you. However, while it’s true that your computer picked the color you see, it did so from the range of possible choices which I specified. I know that it didn’t show you yellow, or green, or brown, because I decided not to include those colors as options. Ultimately, I’m in control of your visual experience, not your computer.
Testing the Theory
Imagine a programmer creates software which doesn’t create art, but instead creates other programs which create art. Even in this scenario, the programming of the secondary software is still controlled by the programmer, since the programmer must specify the realm of possible secondary software the first program is capable of producing. In fact, no matter how many programs of this sort are placed between the programmer and the final output, ultimately that output is still created according to the initial specifications by the programmer of the first program.
Now imagine someone builds a robot which continuously seeks out art and then creates new art based on what it found. Assume this robot is able to consider the infinite aspects of artworks, so the robot’s creator doesn’t have to codify which aspects the robot should consider when examining and creating art. Has the creator managed to remove their taste from this program’s output? Still no, as the creator will be forced to add their own taste regarding what should and should not be considered art. For example, how will it know that Duchamp’s Fountain is art, but the toilets in my house are not art? How will it know that The Mona Lisa is art but the diagrams on my whiteboard are not art? Yes, even just defining “art” inherently reveals your taste (perhaps you disagree that Fountain makes the cut), and a computer can’t create art without a definition of it.
Computers can only do what we tell them to. This implies that humans must use their taste to dictate any art which is created by software. So artists can never be replaced by algorithms, because at the other end of every piece of computer-generated art you’ll find an artist.
Thank you so much for reading.
I make endlessly unique generative music on Generative.fm, which is completely free and open source. You might enjoy it if you like ambient music or you need something to listen to while you concentrate. You can even generate unique recordings for use in your own projects, including commercial ones.
I tweet links which bypass Medium’s paywall for all of my articles, so you might consider following me especially if you’re not a Medium member. Or, you can subscribe to my newsletter to get paywall-free links.